A framework for combining symbolic and connectionist learning with equivalent concept descriptions

We propose a learning formalism combining symbolic and connectionist learning methods. As opposed to previous such work, we require the two methods to use equivalent concept descriptions (use the same mechanism for dividing the concept space when representing learned concepts), thus taking advantage of a combination of two different but related learning biases while still retaining the understandability of the learned concepts. We give two examples of applying the formalism and experimentally evaluate one of them (using decision trees and M-RCE networks). The integration of the two paradigms in our formalism and its application to several domains with different characteristics yielded consistently better classification performance than each of its separate components.